A Dynamic Graph Model of Strategy Learning for Predicting Human Behavior in Repeated Games

We present a model that explains the process of strategy learning by the players in repeated normal-form games. The proposed model is based on a directed weighted graph, which we define and call as the game's dynamic graph. This graph is used as a framework by a learning algorithm that predicts which actions will be chosen by the players during the game and how the players are acting based on their gained experiences and behavioral characteristics. We evaluate the model's performance by applying it to some human-subject datasets and measure the rate of correctly predicted actions. The results show that our model obtains a better average hit-rate compared to that of respective models. We also measure the model's descriptive power (its ability to describe human behavior in the self-play mode) to show that our model, in contrast to the other behavioral models, is able to describe the alternation strategy in the Battle of the sexes game and the cooperating strategy in the Prisoners' dilemma game.